modeling_diff_llama.py

import math
from typing import Optional, Tuple, Union, List, Dict

import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
from transformers import PreTrainedModel, LlamaConfig
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.models.llama.modeling_llama import (
    LlamaRMSNorm,
    LlamaRotaryEmbedding,
    LlamaLinearScalingRotaryEmbedding,
    LlamaDynamicNTKScalingRotaryEmbedding,
    LlamaMLP,
    apply_rotary_pos_emb,
    repeat_kv,
)
from transformers.cache_utils import Cache, DynamicCache, StaticCache


class DiffLLaMAConfig(LlamaConfig):
    """
    Configuration class for the DiffLLaMA model.
    Inherits from LlamaConfig and can be extended with additional parameters.
    """
    model_type = "diff_llama"
    
    def __init__(
        self,
        num_kv_heads: int = 8,
        intermediate_size: int = 3072,
        rope_scaling: Optional[Dict[str, Union[str, float]]] = None,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.num_kv_heads = num_kv_heads
        self.intermediate_size = intermediate_size
        self.rope_scaling = rope_scaling or {"type": "linear", "factor": 1.0}
        # Add any custom configuration parameters here


def init_method(tensor):
    """Initialize tensor with Kaiming uniform initialization."""
    nn.init.kaiming_uniform_(tensor, a=math.sqrt(5))

def lambda_init_fn(depth):
    """Compute lambda initialization value based on layer depth."""
    return 0.8 - 0.6 * math.exp(-0.3 * depth)

class MultiheadDiffAttn(nn.Module):
    def __init__(self, config: DiffLLaMAConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_kv_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.scaling = self.head_dim ** -0.5

        self.rotary_emb = self._init_rope()

        self.lambda_init = lambda_init_fn(layer_idx if layer_idx is not None else 0)
        self.lambda_q1 = nn.Parameter(torch.zeros(self.head_dim, dtype=torch.float32).normal_(mean=0, std=0.1))
        self.lambda_k1 = nn.Parameter(torch.zeros(self.head_dim, dtype=torch.float32).normal_(mean=0, std=0.1))
        self.lambda_q2 = nn.Parameter(torch.zeros(self.head_dim, dtype=torch.float32).normal_(mean=0, std=0.1))
        self.lambda_k2 = nn.Parameter(torch.zeros(self.head_dim, dtype=torch.float32).normal_(mean=0, std=0.1))

        self.subln = nn.LayerNorm(self.num_heads * self.head_dim, elementwise_affine=False)

        self._init_rope()

    def _init_rope(self):
        if not hasattr(self.config, 'rope_scaling') or self.config.rope_scaling is None:
            self.rotary_emb = LlamaRotaryEmbedding(
                self.head_dim,
                max_position_embeddings=self.max_position_embeddings,
                base=self.rope_theta,
            )
        else:
            scaling_type = self.config.rope_scaling.get("type", "linear")
            scaling_factor = self.config.rope_scaling.get("factor", 1.0)
            if scaling_type == "linear":
                self.rotary_emb = LlamaLinearScalingRotaryEmbedding(
                    self.head_dim,
                    max_position_embeddings=self.max_position_embeddings,
                    scaling_factor=scaling_factor,
                    base=self.rope_theta,
                )
            elif scaling_type == "dynamic":
                self.rotary_emb = LlamaDynamicNTKScalingRotaryEmbedding(
                    self.head_dim,
                    max_position_embeddings=self.max_position_embeddings,
                    scaling_factor=scaling_factor,
                    base=self.rope_theta,
                )
            else:
                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        batch_size, seq_length, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(batch_size, seq_length, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(batch_size, seq_length, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value[0].shape[-2]
        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        if past_key_value is not None:
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)

        past_key_value = (key_states, value_states) if use_cache else None

        # Repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(-1, -2))
        attn_weights = attn_weights * self.scaling

        lambda_1 = torch.exp(torch.sum(self.lambda_q1 * self.lambda_k1))
        lambda_2 = torch.exp(torch.sum(self.lambda_q2 * self.lambda_k2))
        lambda_full = lambda_1 - lambda_2 + self.lambda_init

        # Apply differential attention
        attn_weights_diff = attn_weights[:, :, :, :-1] - lambda_full * attn_weights[:, :, :, 1:]
        attn_weights = torch.cat([attn_weights_diff, attn_weights[:, :, :, -1:]], dim=-1)

        if attention_mask is not None:
            # Expand attention_mask
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
            attention_mask = attention_mask.expand(batch_size, self.num_heads, seq_length, attention_mask.size(-1))
            attention_mask = attention_mask.to(dtype=attn_weights.dtype)  # Convert to same dtype as attn_weights
            
            # Use a large negative number instead of negative infinity
            attn_weights = attn_weights + (1.0 - attention_mask) * -10000.0

        attn_weights = F.softmax(attn_weights, dim=-1)

        attn_output = torch.matmul(attn_weights, value_states)
        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, seq_length, self.num_heads * self.head_dim)

        attn_output = self.subln(attn_output)
        attn_output = attn_output * (1 - self.lambda_init)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class DiffLLaMALayer(nn.Module):
    """
    A single layer of the DiffLLaMA model, consisting of multi-head differential attention and a feed-forward network.
    Incorporates gradient checkpointing for memory efficiency.
    """
    def __init__(self, config: DiffLLaMAConfig, layer_idx: int):
        super().__init__()
        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.self_attn = MultiheadDiffAttn(
            config=config,
            layer_idx=layer_idx
        )
        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.mlp = LlamaMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs

class DiffLLaMAModel(PreTrainedModel):
    """
    DiffLLaMAModel is a variant of LLaMA with differential attention mechanisms.
    Incorporates mixed precision training and gradient checkpointing for optimized performance.
    """
    config_class = DiffLLaMAConfig
        
    def __init__(self, config: DiffLLaMAConfig):
        super().__init__(config)
        self.config = config

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([
            DiffLLaMALayer(config, layer_idx=i) for i in range(config.num_hidden_layers)
        ])
        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        
        self.rotary_emb = LlamaRotaryEmbedding(
            dim=config.hidden_size // config.num_attention_heads,
            max_position_embeddings=config.max_position_embeddings,
            base=config.rope_theta,
        )
        
        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()
    
    def forward(
            self,
            input_ids: Optional[torch.LongTensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[Tuple[torch.FloatTensor, torch.FloatTensor]]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            cache_position: Optional[torch.LongTensor] = None,
        ) -> Union[Tuple, BaseModelOutputWithPast]:
        
        """
        Forward pass for the DiffLLaMAModel with performance optimizations.

        Args:
            input_ids: Input token IDs.
            attention_mask: Attention mask.
            position_ids: Position IDs.
            past_key_values: Past key and value tensors for caching.
            inputs_embeds: Input embeddings.
            use_cache: Whether to return present key and value for caching.
            output_attentions: Whether to output attention weights.
            output_hidden_states: Whether to output hidden states.
            return_dict: Whether to return a dict.
            cache_position: Position IDs for caching.

        Returns:
            Model output, either as a tuple or a BaseModelOutputWithPast.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).expand(batch_size, -1)

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        # Position embeddings are handled within each layer; remove pre-computation
        # Removed the following lines:
        # cos, sin = self.rotary_emb(position_ids, seq_len=seq_length)
        # position_embeddings = (cos, sin)

        hidden_states = inputs_embeds

        # Attention mask
        if attention_mask is None:
            attention_mask = torch.ones((batch_size, seq_length), device=hidden_states.device)

        # Initialize lists to store outputs
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_cache = () if use_cache else None

        for idx, layer in enumerate(self.layers):
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = layer(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_values[idx] if past_key_values is not None else None,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
            )

            # Correctly unpack layer_outputs based on the configuration
            hidden_states = layer_outputs[0]
            
            if use_cache:
                present_key_value = layer_outputs[-1]
                next_cache += (present_key_value,)

            if output_attentions:
                self_attn_weights = layer_outputs[1]
                all_self_attns += (self_attn_weights,)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = None
        if use_cache:
            next_cache = (
                next_cache.to_legacy_cache() if isinstance(next_cache, Cache) else next_cache
            )
        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

class DiffLLaMAForCausalLM(PreTrainedModel):
    """
    DiffLLaMA model with a causal language modeling head.
    Incorporates mixed precision training for optimized performance.
    """
    config_class = DiffLLaMAConfig
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: DiffLLaMAConfig):
        super().__init__(config)
        self.model = DiffLLaMAModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        """Return input embeddings."""
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        """Set input embeddings."""
        self.model.set_input_embeddings(value)

    def get_output_embeddings(self):
        """Return output embeddings (language modeling head)."""
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        """Set output embeddings (language modeling head)."""
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        """Set the decoder model."""
        self.model = decoder

    def get_decoder(self):
        """Get the decoder model."""
        return self.model

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[Tuple[torch.FloatTensor, torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        """
        Forward pass for DiffLLaMAForCausalLM with performance optimizations.

        Args:
            input_ids: Input token IDs.
            attention_mask: Attention mask.
            position_ids: Position IDs.
            past_key_values: Past key and value tensors for caching.
            inputs_embeds: Input embeddings.
            labels: Labels for computing the loss.
            use_cache: Whether to return past key and value tensors.
            output_attentions: Whether to output attention weights.
            output_hidden_states: Whether to output hidden states.
            return_dict: Whether to return a dict.
            cache_position: Position IDs for caching.

        Returns:
            CausalLMOutputWithPast or tuple containing loss and logits.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # Get outputs from the model
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = outputs.last_hidden_state if return_dict else outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Compute loss using mixed precision if enabled
            if shift_logits.dtype == torch.float16:
                with torch.cuda.amp.autocast(enabled=False):
                    loss = loss_fct(shift_logits, shift_labels)
            else:
                loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            if use_cache:
                return ((loss, logits) + outputs[1:]) if loss is not None else (logits,) + outputs[1:]
            else:
                return (loss, logits) if loss is not None else (logits,)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "cache_position": kwargs.get("cache_position"),
            }
        )
        return model_inputs